Gated Spectral State Space Image Encoding for Long Sequences

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Transformer and Mamba frameworks face challenges with quadratic computational complexity, increased learning parameters, and training instability, particularly when handling long input sequences, leading to high latency and performance gaps compared to state-of-the-art models.

Innovation Solution

The introduction of a gated spectral state space model (GSSSM) that performs spectral transformations of embedded input image patches, eliminating the need for an initial convolutional neural network layer and using parametrized Gaussian functions to approximate complex matrix-based computations, thereby reducing training parameters and latency while maintaining sub-quadratic computational efficiency.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If transformer models are used for processing long sequences, then comprehensive feature representation is achieved, but computational complexity increases quadratically

Engineering Contradiction:
Improvefeature representation qualityVSAvoidcomputational complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent transforms the state space model parameters into the spectral domain using Fourier or cosine transforms. This parameter transformation allows the model to process sequences with reduced computational complexity while maintaining the ability to capture long-range dependencies, directly resolving the contradiction between comprehensive feature representation and quadratic computational complexity

Inventive Principle:
Principle #35Parameter changes

Solution Approach 2:

The patent replaces the traditional time-domain state space model computations with frequency-domain operations. By substituting sequential matrix multiplications with spectral transforms and element-wise operations, the model achieves linear or near-linear complexity while preserving representational power

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

2Productivity

If Mamba framework is used to reduce computational complexity, then processing efficiency improves, but training stability deteriorates

Engineering Contradiction:
Improveprocessing efficiencyVSAvoidtraining stability
Core Design Contradiction:
ProductivityVSStability of the object's composition

Solution Approach 1:

The patent introduces spectral transforms as an intermediary layer between the input and the state space model computations. This spectral domain acts as a mediator that stabilizes the training process by transforming the optimization landscape, making it more amenable to gradient-based methods while maintaining the efficiency benefits of the Mamba framework

Inventive Principle:
Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent moves the computations from the time domain to the frequency domain, effectively changing the dimension in which operations are performed. This dimensional transformation stabilizes training by decoupling different frequency components, allowing each to be processed independently with more stable gradient updates

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

3Manufacturing precision

If initial convolutional neural network layer is used for image processing, then local feature extraction is improved, but overall model latency increases

Engineering Contradiction:
Improvefeature extraction qualityVSAvoidmodel latency
Core Design Contradiction:
Manufacturing precisionVSLoss of time

Solution Approach 1:

The patent extracts and removes the initial convolutional neural network layer from the traditional vision transformer architecture. By taking out this layer, the model directly processes patch embeddings through the spectral state space model, eliminating the latency introduced by convolutional operations while the spectral transforms compensate for local feature extraction capabilities

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The spectral state space model serves multiple functions simultaneously: it performs both the local feature extraction traditionally done by convolutions and the global sequence modeling. This multi-functionality eliminates the need for separate convolutional layers, reducing latency while maintaining feature extraction quality

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP4687118A1Gated spectral state space model for image encoding
Publication Date: 2026.02.04 MICROSOFT TECHNOLOGY LICENSING LLC
  • EP4687118A1 patent drawingFigure 1
  • EP4687118A1 patent drawingFigure 2
  • EP4687118A1 patent drawingFigure 3

AI summary

A system may generate embedded subsets by projecting each subset of the subsets into a vector space to generate a corresponding embedded subset. A system may encode the embedded subsets into an encoded image using a dataset encoder including a gated spectral state space model, the gated spectral state space model being a gated neural network that includes a spectral state space model, the spectral state space model being a state space model that represents features of the input dataset using at least a spectral transformation of each embedded subset of the embedded subsets. A system may predict a classification for the input dataset using the encoded image.